A REAL-TIME VISION-BASED PARKING SLOT OCCUPANCY DETECTION USING YOLOV8 UNDER VARYING ILLUMINATION CONDITIONS

Abstract

Due to rapid urbanization efficient parking management has become a critical challenge for cities leading to parking congestions and adverse impact of it on environment. Computer vision-based parking system is a scalable and cost-effective solution to manage smart cities. This paper presents a real time, computer vision based YOLOv8 Object Detection model for intelligent parking system which is designed to operate robustly under different environment and lighting conditions. The proposed model was initially trained and validated using YOLOv8 on publicly available PK_Lot data subset containing 6998 annotated images under different weather conditions. To enhance real world applicability, the evaluation was conducted on continuous CCTV video captured from different parking environment. Manually annotated frames are extracted from real word CCTV video to address multiple real word parking challenges such as motion blur, elimination glare, different camera angles, partial occlusions and improper parking layout. As the results on real world data set are not satisfactory so the combined data set with PK_Lot and annotated frames from real CCTV recordings included for enhanced generalization capabilities. A total of 8798 annotated frames are extracted and 879 frames are used for structural testing. The experimental results obtained from enhanced dataset gives high detection accuracy of 99.49%, precision of 99.69%, recall of 99.47% and F1 score of 99.58% demonstrating reliable performance under different practical deployment conditions. The result confirmed the robustness of the proposed framework against real -world conditions, making proposed system suitable for scalable intelligent parking management system for smart city.